Skip to content

GlassAlpha

GlassAlpha makes deterministic, regulator-ready audit reports for tabular ML models. It's an open-source (Apache 2.0) toolkit for reproducible, audit-ready model documentation (HTML/PDF).

Note: GlassAlpha is currently in beta (v0.2.0). Core functionality is stable with 1000+ passing tests and comprehensive documentation. Breaking API changes may occur before v1.0. First stable release expected Q1 2025.

How it works

Single command: glassalpha audit --config your_config.yaml --output audit.html

graph TB
    A[Your Model + Data]
    A --> B[GlassAlpha CLI]

    B --> Pipeline

    subgraph Pipeline[Audit Pipeline]
        D[Performance Analysis]
        E[Fairness Testing]
        F[Explainability]
        G[Calibration Testing]
    end

    D --> H
    E --> H
    F --> H
    G --> H

    H[Professional PDF]
    H --> I[Evidence Pack SHA256]
    I --> J[Regulatory Submission]

    style A fill:#e1f5ff
    style Pipeline fill:#f0f0f0
    style H fill:#d4edda
    style I fill:#fff3cd
    style J fill:#f8d7da

New users:

Key features:

Run your first audit in 60 seconds

# 1. Install with pipx (30 seconds)
pipx install glassalpha

# 2. Generate a quickstart project
glassalpha quickstart

# 3. Run audit (choose your path - both work!)
cd my-audit-project
python run_audit.py          # Option A: Uses generated script (great for learning)
# OR
glassalpha audit             # Option B: Direct CLI (great for CI/CD)

# 4. Done! Open your professional report (~2-3 seconds total)
open reports/audit_report.html  # macOS
# xdg-open reports/audit_report.html  # Linux
# start reports/audit_report.html  # Windows

Or install from source:

git clone https://github.com/GlassAlpha/glassalpha
cd glassalpha && pip install -e .
glassalpha audit --config configs/german_credit_simple.yaml --output audit.html --fast

Result: A professional audit report with model performance, fairness analysis, feature importance, individual explanations, reason codes, preprocessing verification, and complete audit trail.

See detailed quickstart guide →

What you get

Every GlassAlpha audit includes:

  • ✅ Model performance metrics (accuracy, precision, recall, F1, AUC)
  • ✅ Group fairness with confidence intervals (demographic parity, equal opportunity, predictive parity)
  • ✅ Intersectional fairness (bias at demographic intersections like gender×race)
  • ✅ Individual fairness (consistency testing, matched pairs, counterfactual analysis)
  • ✅ Dataset bias detection (proxy correlations, distribution drift, sampling power)
  • ✅ Calibration analysis with confidence intervals (ECE, Brier score, calibration curves)
  • ✅ Robustness testing (adversarial perturbations, demographic shift simulation)
  • ✅ Feature importance (coefficient-based for linear models, TreeSHAP for gradient boosting)
  • ✅ Individual explanations (SHAP values for specific predictions)
  • ✅ Preprocessing verification (production artifact validation with dual hash system)
  • ✅ Complete audit trail (reproducibility manifest with all seeds, versions, and hashes)

See complete audit contents →

Additional commands available:

  • Reason codes: Generate ECOA-compliant adverse action notices via glassalpha reasons
  • Recourse recommendations: Generate counterfactual explanations via glassalpha recourse

See CLI reference for all commands →

Current limitations

GlassAlpha is beta software. Current limitations:

  • Tabular data only - No text, image, or time series support yet
  • Classification focus - Limited regression support
  • Cross-platform PDF differences - Byte-identical PDFs within same platform+Python version, but not across Ubuntu/macOS due to font rendering and library differences. Regulators can reproduce audits on their infrastructure. See determinism guide for details
  • Single machine - No distributed computing support

See complete limitations and roadmap →

Why choose GlassAlpha?

GlassAlpha is the only OSS tool that combines professional audit reports, easy custom data support, and complete regulatory compliance.

  • Deterministic outputs - Byte-identical HTML reports on same platform+Python+config
  • Complete lineage - Git SHA, config hash, data hash, seeds recorded
  • Professional formatting - Publication-quality reports with visualizations
  • No external dependencies - Runs completely offline
  • Single command - glassalpha audit handles everything

See detailed comparison and features →

Use cases

Documentation

License & trademark

  • License: Apache 2.0 - See LICENSE
  • Trademark: While GlassAlpha's code is open source, the brand is not. We respectfully request that our name and logo not be used in confusing or misleading ways. See TRADEMARK.

Support

Built for teams who need reproducible, regulator-ready ML audit reports.